Embed to Control

This is a pytorch implementation of the paper "Embed to Control: A Locally Linear Latent Dynamics Model for Control from Raw Images", NIPS, 2015.

Note: This is not and official implementation.

Installing

First, clone the repository:

git clone https://github.com/tungnd1705/E2C-pytorch.git

Install the dependencies as listed in env.yml and activate the environment

conda env create -f env.yml

conda activate e2c

Then install the patch version of gym in order to sample the pendulum data

cd gym

python setup.py install

Simulate training data

Currently the code supports simulating 3 environments: planar, pendulum and cartpole.

In order to generate data, simply run python sample_{env_name}_data.py --sample_size={sample_size}.

Note: the sample size is equal to the total number of training and test data

For the planar task, we base on this implementation and modify for our needs.

Training

Run the train_e2c.py with your own settings. Example:

python train_e2c.py 
    --env=planar 
    --propor=3/4 
    --batch_size=128 
    --lr=0.0001 
    --lam=0.25 
    --num_iter=5000 
    --iter_save=1000

You can visualize the training process by running tensorboard --logdir=logs.

Citation

If you find E2C useful in your research, please consider citing:

@inproceedings{watter2015embed,
  title={Embed to control: A locally linear latent dynamics model for control from raw images},
  author={Watter, Manuel and Springenberg, Jost and Boedecker, Joschka and Riedmiller, Martin},
  booktitle={Advances in neural information processing systems},
  pages={2746--2754},
  year={2015}
}